Deep Learning-Based Building Carbon Emission Data Processing Method and Computer System

Abstract

This invention provides a deep learning-based method and computer system for processing building carbon emission data. By acquiring a set of building carbon emission data, feature association mining is performed to generate a correlation strength tensor. The first dimension of the correlation strength tensor corresponds to the functional unit, the second dimension to the time series position, and the third dimension to the feature association type. Tensor elements represent the degree of correlation between the carbon emission features of the corresponding functional unit at the corresponding time series position and the carbon emission features of other functional units. A time-varying dependency network is constructed based on the correlation strength tensor to obtain the time-varying dependency network parameters. Multi-scale evolution trajectories are tracked using the time-varying dependency network parameters to generate a set of multi-scale evolution trajectories. Attention mechanism key path identification is performed using the multi-scale evolution trajectory set to output the carbon emission data processing results. This invention can improve the effectiveness and reliability of the overall carbon emission data processing results.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and data processing technology, and in particular to a deep learning-based method and computer system for processing building carbon emission data. Background Technology

[0002] With the increasing demand for carbon emission reduction in the construction industry, research on building energy conservation and carbon management has become a key focus. Analyzing carbon emission monitoring data and related impact data from building functional systems helps identify critical pathways influencing carbon emission characteristics, providing data support for optimizing low-carbon building operations. Currently, building carbon emission data processing uses two-dimensional matrices to analyze the relationships between systems. Static networks are constructed to describe the dependencies between functional units, tracking the evolution of carbon emission characteristics and relying on preset thresholds or manual rules to filter critical impact paths. However, existing technologies cannot distinguish between different types of correlation characteristics, and static networks fix the connections between nodes, making it difficult to capture changes in correlation strength over time, thus affecting the effectiveness and reliability of carbon emission data processing results. Summary of the Invention

[0003] In view of this, the present invention provides a method and computer system for processing building carbon emission data based on deep learning. The technical solution of the embodiments of the present invention is implemented as follows:

[0004] On one hand, embodiments of the present invention provide a deep learning-based method for processing building carbon emission data, comprising: acquiring a building carbon emission dataset, the building carbon emission dataset including a carbon emission monitoring sequence and a related impact sequence of a building functional system cluster in a continuous time series, the building functional system cluster being a combination of functional units that generate carbon emissions during building operation, and the related impact sequence being a sequence of internal and external environmental parameters that affect the changes in the carbon emission characteristics of functional units over time; performing feature association mining on the building carbon emission dataset to generate an association strength tensor, the first dimension of the association strength tensor corresponding to the functional unit, the second dimension corresponding to the time series position, and the third dimension corresponding to the feature association type, the tensor elements representing the degree of association between the carbon emission characteristics of the corresponding functional unit at the corresponding time series position and the carbon emission characteristics of other functional units; constructing a time-varying dependency network based on the association strength tensor to obtain time-varying dependency network parameters; tracking multi-scale evolution trajectories through the time-varying dependency network parameters to generate a multi-scale evolution trajectory set; performing attention mechanism critical path identification through the multi-scale evolution trajectory set, and outputting carbon emission data processing results.

[0005] On the other hand, embodiments of the present invention provide a computer system including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the program to implement the steps in the above-described method.

[0006] This invention provides a deep learning-based method for processing building carbon emission data. It acquires a building carbon emission dataset, merges it, performs feature association mining to generate an association strength tensor, constructs a time-varying dependency network to obtain its parameters, tracks multi-scale evolution trajectories to generate a multi-scale evolution trajectory set, and then identifies key paths through an attention mechanism to output the processing results. This invention extends association analysis from two-dimensional to three-dimensional using the association strength tensor, enabling the representation of multiple types of association features. Combined with the time-varying dependency network, it allows dependency modeling to evolve over time. Multi-scale evolution trajectory tracking captures short-term, medium-term, and long-term evolutionary patterns in a hierarchical manner. The attention mechanism-driven key path identification adaptively selects path sequences dominated by contribution, effectively improving the accuracy of association feature characterization, the reliability of dependency modeling, the comprehensiveness of evolutionary pattern capture, and the accuracy of key path identification in the building carbon emission data processing process, thereby enhancing the overall effectiveness and reliability of the carbon emission data processing results. Attached Figure Description

[0007] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present invention and, together with the specification, serve to explain the technical solutions of the present invention.

[0008] Figure 1 This is a schematic diagram illustrating the implementation process of a deep learning-based building carbon emission data processing method provided in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of the hardware entity of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] This invention provides a deep learning-based method for processing building carbon emission data, which can be executed by a computer system's processor. The computer system can refer to devices with data processing capabilities, such as servers or personal PCs.

[0012] Figure 1 This is a schematic diagram illustrating the implementation process of a deep learning-based building carbon emission data processing method provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes the following steps S100~S500 and their respective sub-steps:

[0013] Step S100: Obtain a building carbon emission data set, which includes a carbon emission monitoring sequence and related impact sequence of a building functional system cluster in a continuous time series. The building functional system cluster is a combination of functional units that generate carbon emissions during building operation, and the related impact sequence is a sequence of internal and external environmental parameters that affect the changes in the carbon emission characteristics of the functional units over time.

[0014] A building carbon emissions dataset is a collection and integration of data related to building carbon emissions. It includes continuous time-series carbon emission monitoring sequences and associated impact sequences for building functional system clusters. A building functional system cluster refers to a combination of functional units that generate carbon emissions during building operation. These units include, for example, air conditioning systems, lighting systems, and elevator systems, which consume energy and generate corresponding carbon emissions during operation. The carbon emission monitoring sequence is a data sequence obtained by monitoring the carbon emissions of these functional units over a continuous time series. For example, carbon emission sensors installed on each functional unit periodically collect the carbon emissions of each unit, thus forming the carbon emission monitoring sequence. The associated impact sequence is a sequence of internal and external environmental parameters that affect the changes in the carbon emission characteristics of functional units over time. External environmental parameters include outdoor temperature, humidity, and light intensity, while internal environmental parameters include indoor occupancy density and equipment usage frequency. These parameters all affect the carbon emission characteristics of functional units. For example, in hot summers, increased outdoor temperatures lead to increased energy consumption of air conditioning systems, thereby increasing their carbon emissions.

[0015] To obtain building carbon emission datasets, various data acquisition devices and methods can be employed. For carbon emission monitoring sequences, specialized carbon emission sensors can be installed on each functional unit. These sensors can monitor the energy consumption of the functional units in real time and convert energy consumption into carbon emissions based on a pre-set carbon emission calculation model. For related impact sequences, different types of sensors can be used to collect internal and external environmental parameters. For example, temperature sensors can be used to collect outdoor and indoor temperature data, light sensors can be used to collect light intensity data, and occupancy counters can be used to count indoor occupancy density.

[0016] Step S200: Perform feature association mining on the building carbon emission dataset to generate an association strength tensor. The first dimension of the association strength tensor corresponds to the functional unit, the second dimension corresponds to the time series position, and the third dimension corresponds to the feature association type. The tensor elements represent the degree of association between the carbon emission characteristics of the corresponding functional unit at the corresponding time series position and the carbon emission characteristics of other functional units.

[0017] Feature association mining is the process of analyzing building carbon emission datasets to uncover the relationships between carbon emission characteristics of different functional units and between carbon emission characteristics and associated impact sequences. The association strength tensor is a three-dimensional tensor used to represent these relationships. Its first dimension corresponds to the functional unit, meaning each functional unit occupies a position in the tensor; the second dimension corresponds to the time series position, reflecting the association at different time points; and the third dimension corresponds to the type of feature association, which can include different types of associations such as linear and non-linear associations. Tensor elements characterize the degree of association between the carbon emission characteristics of a corresponding functional unit at a corresponding time series position and the carbon emission characteristics of other functional units. The higher the degree of association, the stronger the correlation between the carbon emission characteristics of the two functional units.

[0018] As one implementation method, step S200 involves performing feature association mining on the building carbon emission dataset to generate an association strength tensor, which can be specifically implemented as the following steps S210~S260:

[0019] Step S210: Separate the carbon emission monitoring subsequence and associated impact subsequence of each functional unit from the building carbon emission data set to obtain the functional unit data sequence set. Each element in the functional unit data sequence set is a combination sequence of the carbon emission monitoring subsequence and associated impact subsequence corresponding to a functional unit.

[0020] The purpose of separating the carbon emission monitoring subsequences and related impact subsequences of each functional unit from the building carbon emission dataset is to process the data of different functional units independently. The carbon emission monitoring subsequence is the carbon emission monitoring data of each functional unit within a continuous time series, and the related impact subsequence is the related impact sequence data corresponding to that functional unit. Combining these two subsequences yields the combined sequence for each functional unit, and the combined sequences of all functional units constitute the functional unit data sequence set. For example, for a building's air conditioning system, its carbon emission monitoring subsequence records the carbon emissions of the air conditioning system at different time points, and the related impact subsequence records the internal and external environmental parameters related to the carbon emissions of the air conditioning system, such as outdoor temperature and indoor occupancy density. Combining these two subsequences yields the combined sequence for the air conditioning system. To achieve data separation, filtering and extraction can be performed based on data identification information. During the data acquisition process, each functional unit's data will have a corresponding identifier, which can be used to distinguish the data of different functional units. Then, the carbon emission monitoring subsequence and related impact subsequence of each functional unit are extracted separately and combined into a combined sequence.

[0021] Step S220: Perform feature coupling processing on the combined sequences in the functional unit data sequence set to generate coupled feature sequences. Feature coupling processing interweaves the carbon emission monitoring subsequence and the associated impact subsequence according to the feature dimension of the time series position, so that the feature vector of each time series position simultaneously contains carbon emission monitoring features and associated impact features.

[0022] Feature coupling processing involves further processing the combined sequences in the functional unit data sequence set, fusing the carbon emission monitoring subsequence with the related impact subsequence. By interleaving the two subsequences along the feature dimension according to their time series positions, the feature vector at each time series position simultaneously contains both carbon emission monitoring features and related impact features. This provides a more comprehensive reflection of the functional unit's status at each time point. For example, at a certain time point, the feature vector contains both the functional unit's carbon emissions and related impact features such as outdoor temperature and indoor population density at that time.

[0023] As one implementation method, step S220 involves performing feature coupling processing on the combined sequences in the functional unit data sequence set to generate coupled feature sequences. Specifically, this can be implemented as the following steps S221~S225:

[0024] Step S221: Obtain the carbon emission monitoring subsequence and the associated impact subsequence in the combined sequence, and determine the time series length and feature dimension of the two subsequences. The time series length is the number of positions in the continuous time series, and the feature dimension is the number of feature parameters contained in each time series position.

[0025] Before performing feature coupling processing, it is necessary to first obtain the carbon emission monitoring subsequence and the associated impact subsequence from the combined sequence, and determine their time series length and feature dimensions. Time series length refers to the number of locations in a continuous time series. For example, if the data is collected hourly for 24 hours, then the time series length is 24. Feature dimension refers to the number of feature parameters contained at each time series location. For example, each time point in the carbon emission monitoring subsequence may contain multiple feature parameters such as carbon emissions and energy consumption, while each time point in the associated impact subsequence may contain multiple feature parameters such as outdoor temperature, humidity, and light intensity. The carbon emission monitoring subsequence and the associated impact subsequence can be extracted by parsing the combined sequence, and then their time series length and feature dimensions can be calculated. For example, data processing software can be used to analyze the combined sequence, traversing each element in the sequence, calculating the time series length, and recording the number of feature parameters at each time point.

[0026] Step S222: Perform time series alignment verification on the carbon emission monitoring subsequence and the associated impact subsequence to ensure that the time series lengths of the two subsequences are consistent and the time series positions correspond one-to-one.

[0027] Time series alignment verification is the process of ensuring that the carbon emission monitoring subsequence and the associated impact subsequence are synchronized in time. Since the two subsequences may be collected by different sensors, and the collection times may differ, alignment verification is required to ensure that the time series lengths of the two subsequences are consistent and that their time series positions correspond one-to-one. This ensures that the features of the two subsequences can be accurately interleaved in subsequent feature coupling processing.

[0028] Interpolation or deletion methods can be used for time series alignment verification. If the time series lengths of two subsequences are inconsistent, missing time point data can be inserted into the shorter subsequence through interpolation, or redundant time point data can be deleted from the longer subsequence to make the time series lengths of the two subsequences consistent.

[0029] Step S223: Perform dimensional interleaving between the feature vector of the carbon emission monitoring subsequence at each time series position and the feature vector of the associated influence subsequence at the corresponding time series position. Dimensional interleaving is achieved by alternating the two feature vectors according to the order of feature parameters to generate interleaved feature vectors.

[0030] Dimensional interleaving is a crucial step in fusing the feature vectors of carbon emission monitoring subsequences and related impact subsequences. Interleaved feature vectors are generated by alternating the two feature vectors according to their feature parameters. For example, assuming the feature vector of the carbon emission monitoring subsequence at a certain time point is [a1,a2,a3], and the feature vector of the related impact subsequence at the same time point is [b1,b2,b3], then the interleaved feature vector is [a1,b1,a2,b2,a3,b3]. This allows for a thorough fusion of the feature information from the two subsequences, resulting in feature vectors at each time point containing richer information.

[0031] Step S224: Enhance the features of the interleaved feature vector by inserting differential features of adjacent time series positions into the interleaved feature vector. The differential features are the difference between the interleaved feature vectors of the current time series position and the previous time series position.

[0032] Feature enhancement aims to further enrich the information in interleaved feature vectors and improve the accuracy of association mining. By inserting difference features from adjacent time series positions into the interleaved feature vectors, the changes in the feature vectors over time can be captured. The difference feature is the difference between the interleaved feature vectors at the current time series position and the previous time series position. For example, assuming the interleaved feature vector at the current time point is [x1, x2, x3] and the interleaved feature vector at the previous time point is [y1, y2, y3], then the difference feature is [x1-y1, x2-y2, x3-y3]. Inserting the difference feature into the interleaved feature vector at the current time point ensures that the feature vector contains not only the feature information at the current time point but also information about feature changes.

[0033] By traversing the interleaved feature vectors, the difference features between adjacent time points can be calculated, and these difference features can be inserted into the interleaved feature vector of the current time point. In the implementation, since the first time point has no previous time point, its difference features can be set to a zero vector.

[0034] Step S225: Arrange the enhanced interleaved feature vectors of all time series positions in time series order to obtain the coupled feature sequence. The time series length of the coupled feature sequence is the same as that of the original combined sequence, and the feature dimension is a preset multiple of the original feature dimension.

[0035] The coupled feature sequence is obtained by arranging the enhanced interleaved feature vectors of all time series positions in chronological order. The time series length of the coupled feature sequence is consistent with that of the original combined sequence, ensuring temporal continuity. The feature dimension is a preset multiple of the original feature dimension. This is because during the feature coupling process, not only are the features of the carbon emission monitoring subsequence and the associated impact subsequence interleaved, but differential features are also inserted, thus expanding the feature dimension. The preset multiple can be set according to specific needs, for example, it can be set to 2 times or 3 times. The coupled feature sequence can also be obtained by storing the enhanced interleaved feature vectors in an array and sorting the array according to the chronological order.

[0036] Step S230: Construct an association mining network. The association mining network includes a sequence encoding layer, a cross-association layer, and a tensor output layer. The sequence encoding layer is used to extract time series features from the coupled feature sequences. The cross-association layer is used to calculate the association between the coupled feature sequences of different functional units. The tensor output layer is used to integrate the association results to generate a tensor.

[0037] The association mining network is a deep learning network used to mine the correlations between carbon emission features of functional units. It consists of a sequence encoding layer, a cross-correlation layer, and a tensor output layer. The sequence encoding layer extracts time-series features from the coupled feature sequences, transforming them into more representative encoded feature sequences. The cross-correlation layer calculates the correlations between the coupled feature sequences of different functional units, identifying these correlations by comparing their encoded feature sequences. The tensor output layer integrates the correlation results, converting the correlation scores calculated by the cross-correlation layer into a correlation strength tensor.

[0038] The sequence encoding layer can be implemented using stacked gated recurrent units (GRUs). GRUs are recurrent neural networks that can effectively process time series data and extract time series features. The cross-association layer can employ a multi-head cross-attention mechanism, which can calculate the association score of the feature sequences encoded by different functional units at each time series position. The tensor output layer can be a three-dimensional convolutional layer, which integrates the association results to generate a tensor by performing convolution operations on the association scores.

[0039] Step S240: Input the set of functional unit data sequences into the sequence encoding layer of the association mining network, and encode the coupled feature sequences of each functional unit through stacked gated recurrent units to obtain the functional unit encoded feature sequence, which contains the time-dependent features of the coupled feature sequence.

[0040] The set of functional unit data sequences is input into the sequence encoding layer of the association mining network. Stacked gated recurrent units (GRUs) are used to encode the coupled feature sequences of each functional unit. A GRU is a special type of recurrent neural network with a gating mechanism, enabling it to process time-series data and capture time-dependent features. By stacking multiple GRU units, the network's expressive power is enhanced, and the time-dependent features of the coupled feature sequences are extracted more accurately. During encoding, the GRU units update the hidden state at the current time point based on the input at the current time point and the hidden state at the previous time point. Through continuous iteration, the hidden state at each time point is obtained. Arranging these hidden states in chronological order yields the encoded feature sequence of the functional unit. This encoded feature sequence contains the time-dependent features of the coupled feature sequences, providing a more comprehensive reflection of the carbon emission characteristics of the functional unit over time.

[0041] Step S250: Input the functional unit encoded feature sequence into the cross-association layer of the association mining network, and calculate the association score of any two functional unit encoded feature sequences at each time series position through the multi-head cross-attention mechanism. The association score is obtained by weighted combination of the attention weight matrix and the encoded feature vector.

[0042] Multi-head cross-attention is a method for calculating the association between different sequences, capturing association information at multiple scales. Functional unit encoded feature sequences are input into the cross-association layer of an association mining network. The multi-head cross-attention mechanism calculates the association score between any two functional unit encoded feature sequences at each time series position. The association score is obtained by a weighted combination of the attention weight matrix and the encoded feature vector. The attention weight matrix reflects the degree of association between the encoded feature sequences of different functional units, and the association score at each time point is obtained by weighting the encoded feature vectors.

[0043] As one implementation method, step S250 involves inputting the functional unit encoded feature sequence into the cross-association layer of the association mining network, and calculating the association score of any two functional unit encoded feature sequences at each time series position through a multi-head cross-attention mechanism. Specifically, this can be implemented as the following steps S251~S258:

[0044] Step S251: Select any two functional unit encoding feature sequences from the functional unit encoding feature sequences as target encoding pairs to obtain the first encoding sequence and the second encoding sequence.

[0045] The purpose of selecting any two functional unit encoding feature sequences from the functional unit encoding feature sequences is to calculate the association score between these two functional units. These two encoding feature sequences are named the first encoding sequence and the second encoding sequence, respectively, for subsequent processing. In practice, this can be achieved by traversing the functional unit encoding feature sequences and sequentially selecting different functional unit pairs for calculation.

[0046] Step S252: Input the first encoding sequence and the second encoding sequence into the query vector generation module and the key-value vector generation module of the multi-head cross-attention mechanism, respectively, to generate a query vector sequence, a key vector sequence and a value vector sequence. The query vector sequence is obtained by linear transformation of the first encoding sequence, and the key vector sequence and the value vector sequence are obtained by different linear transformations of the second encoding sequence.

[0047] The first and second encoded sequences are input into the query vector generation module and key-value vector generation module of the multi-head cross-attention mechanism, respectively. Linear transformations are then used to generate a query vector sequence, a key vector sequence, and a value vector sequence. The query vector sequence, obtained from the first encoded sequence through a linear transformation, is used to find relevant information in the key vector sequence. The key vector sequence and value vector sequence are obtained from the second encoded sequence through different linear transformations. The key vector sequence is used for matching with the query vector sequence, and the value vector sequence is used for weighted summation based on the matching results.

[0048] Linear transformations can be achieved through matrix multiplication. For example, assuming the first encoded sequence is X and the linear transformation matrix of the query vector generation module is Wq, then the query vector sequence Q = X × Wq. Similarly, for the second encoded sequence Y, the key vector sequence K = Y × Wk and the value vector sequence V = Y × Wv, where Wk and Wv are different linear transformation matrices.

[0049] Step S253: Split the query vector sequence, key vector sequence, and value vector sequence according to the number of attention heads to obtain multiple head components. Each head component contains the corresponding query sub-vector sequence, key sub-vector sequence, and value sub-vector sequence.

[0050] The multi-head cross-attention mechanism captures association information at different scales using multiple attention heads. The query vector sequence, key vector sequence, and value vector sequence are split dimensionally according to the number of attention heads, resulting in multiple head components. Each head component contains a corresponding query sub-vector sequence, key sub-vector sequence, and value sub-vector sequence. This allows association scores to be calculated independently in different subspaces, improving the accuracy of association mining.

[0051] The query vector sequence, key vector sequence, and value vector sequence can be partitioned dimensionally, dividing each vector sequence into multiple sub-vector sequences, with each sub-vector sequence corresponding to an attention head. For example, assuming the query vector sequence has a dimension of D and the number of attention heads is H, then the dimension of each query sub-vector sequence is D / H.

[0052] Step S254: Perform a dot product operation on the query sub-vector sequence and the key sub-vector sequence of each head component to obtain the original association matrix. The elements of the original association matrix are the dot product results of the query sub-vector and the key vector.

[0053] The original association matrix is ​​obtained by performing a dot product operation on the query subvector sequence and the key vector sequence for each head component. The dot product operation is a common vector operation; by taking the dot product of query subvectors and key vectors, we can obtain the similarity between them. The elements of the original association matrix are the dot products of query subvectors and key vectors, with each row corresponding to a query subvector and each column corresponding to a key vector. By calculating the original association matrix, we can obtain the degree of association between each query subvector and all key vectors.

[0054] In implementation, the dot product operation can be achieved through matrix multiplication. Assuming the query sub-vector sequence is Q' and the key sub-vector sequence is K', then the original association matrix A = Q' × K'. T , where K' T It is the transpose of the key vector sequence.

[0055] Step S255: Scale the original correlation matrix by dividing each element by the square root of the key vector dimension to obtain the scaled correlation matrix.

[0056] The purpose of scaling the original incidence matrix is ​​to avoid excessively large dot product results, which can lead to vanishing or exploding gradients. By dividing each element of the original incidence matrix by the square root of the key vector dimension, the element values ​​can be controlled within a reasonable range. The scaled incidence matrix is ​​the incidence matrix after scaling, resulting in more stable element values, which is beneficial for subsequent calculations. In practice, the scaled incidence matrix can be obtained by iterating through the original incidence matrix and dividing each element by the square root of the key vector dimension.

[0057] Step S256: Activate the scaling correlation matrix to obtain the attention weight matrix. The elements of the attention weight matrix represent the correlation weights between the query sub-vector sequence and the key sub-vector sequence.

[0058] The purpose of activation processing on the scaled association matrix is ​​to transform the element values ​​of the association matrix into a probability distribution, resulting in the attention weight matrix. Activation processing can employ the softmax function, which converts the input vector into a probability distribution such that each element of the vector is between 0 and 1, and the sum of all element values ​​is 1. The elements of the attention weight matrix represent the association weights between the query sub-vector sequence and the key sub-vector sequence; the larger the weight value, the stronger the association between the query sub-vector and the corresponding key vector. In implementation, the attention weight matrix can be obtained by applying the softmax function to each row of the scaled association matrix. For example, assuming the scaled association matrix is ​​S, the attention weight matrix W = softmax(S).

[0059] Step S257: The attention weight matrix and the value vector sequence are weighted and summed to obtain the head-related feature sequence of each head component.

[0060] The attention weight matrix and the value vector sequence are weighted and summed to obtain the head association feature sequence for each head component. The weighted summation is performed by weighting the value vector sequence according to the element values ​​of the attention weight matrix to obtain the association features corresponding to each query vector. The head association feature sequence is the association result for each head component, reflecting the association information between the query vector sequence and the value vector sequence under that head component. In implementation, weighted summation can be achieved through matrix multiplication. Assuming the attention weight matrix is ​​W and the value vector sequence is V', then the head association feature sequence H = W × V'.

[0061] Step S258: Concatenate the head association feature sequences of all head components in terms of dimensions, and integrate them through a linear transformation layer to obtain the association score of the target encoding pair at each time series position.

[0062] The head association feature sequences of all head components are concatenated dimensionally to integrate the association information of multiple head components. Then, a linear transformation layer further processes the concatenated feature sequence to obtain the association score of the target encoding pair at each time series position. The linear transformation layer can be implemented using matrix multiplication, which converts the concatenated feature sequence into a scalar value as the association score. In the implementation, the head association feature sequences of all head components are concatenated dimensionally to obtain a concatenated feature sequence. Then, matrix multiplication is performed between the concatenated feature sequence and the weight matrix of the linear transformation layer to obtain the association score.

[0063] Step S260: Arrange the association scores according to the functional unit dimension, time series position dimension, and attention head dimension, and input them into the tensor output layer of the association mining network. The features are integrated through a three-dimensional convolutional layer to obtain the association strength tensor.

[0064] The association scores are arranged according to the functional unit dimension, time series position dimension, and attention head dimension to form a three-dimensional data structure. This three-dimensional data is then input into the tensor output layer of the association mining network, where features are integrated through a three-dimensional convolutional layer. The three-dimensional convolutional layer is used to process three-dimensional data, performing convolution operations on the input data to extract feature information. Through the processing of the three-dimensional convolutional layer, the association scores can be integrated to obtain the association strength tensor. In the implementation process, the association scores are first arranged according to the functional unit dimension, time series position dimension, and attention head dimension to form a three-dimensional tensor. Then, this three-dimensional tensor is input into the three-dimensional convolutional layer, setting parameters such as the kernel size and stride, and performing a convolution operation. Finally, the association strength tensor is obtained.

[0065] Step S300: Construct a time-varying dependency network based on the correlation strength tensor to obtain the time-varying dependency network parameters. The time-varying dependency network uses functional units as network nodes and the correlation degree in the correlation strength tensor as the time-varying edge weights between nodes. The time-varying dependency network parameters include the time series of node feature vectors of each network node and the evolution gradient of the time-varying edge weights between nodes.

[0066] Time-varying dependency networks (TDNs) are network models used to describe the time-varying dependencies between functional units. Functional units are treated as network nodes, and the degree of association in the association strength tensor is used as the time-varying edge weights between nodes. By constructing a TDN, the changes in the relationships between different functional units over time can be more intuitively displayed. The parameters of a TDN include the time series of the node feature vectors for each network node and the evolution gradient of the time-varying edge weights between nodes. The time series of the node feature vectors reflects the characteristic information of each node at different time points, and the evolution gradient of the time-varying edge weights reflects the changes in the edge weights over time.

[0067] As one implementation method, step S300, constructing a time-varying dependency network based on the correlation strength tensor to obtain the time-varying dependency network parameters, can be specifically implemented as the following steps S310~S380:

[0068] Step S310: Use the functional unit corresponding to the first dimension of the correlation strength tensor as the initial node set of the time-varying dependency network. Each node in the initial node set corresponds to a functional unit.

[0069] The functional units corresponding to the first dimension of the association strength tensor are used as the initial node set for the time-varying dependency network because the first dimension of the association strength tensor corresponds to different functional units. Each functional unit is treated as a node to construct the initial node set of the time-varying dependency network. In this way, each node represents a specific functional unit in the network, facilitating subsequent analysis of the association relationships between functional units.

[0070] Step S320: Extract the correlation strength matrix for each time series position from the correlation strength tensor. The correlation strength matrix is ​​a two-dimensional matrix formed by the first and third dimensions of the correlation strength tensor when the second dimension is fixed.

[0071] The purpose of extracting the correlation strength matrix for each time series position from the correlation strength tensor is to analyze the correlation relationships between functional units at different time points. The correlation strength matrix is ​​a two-dimensional matrix formed by the first and third dimensions of the correlation strength tensor with the second dimension fixed, reflecting the degree of correlation between different functional units at a given time point. By extracting the correlation strength matrix for each time series position, correlation information at different time points can be obtained. This can be achieved by iterating through the second dimension of the correlation strength tensor, fixing each time point, and extracting the two-dimensional matrix formed by the first and third dimensions to obtain the correlation strength matrix for each time point.

[0072] Step S330: Determine the connection relationship between nodes at each time series position based on the correlation strength matrix. For any two nodes, if the corresponding element value in the correlation strength matrix is ​​greater than the preset correlation threshold, then establish a temporary edge connection between the nodes at that time series position, with the edge weight being the corresponding element value.

[0073] The connection relationships between nodes at each time series position are determined based on the association strength matrix, with the aim of constructing the edge structure of a time-varying dependency network. For any two nodes, if the corresponding element value in the association strength matrix is ​​greater than a preset association threshold, it indicates a strong association between the two nodes. A temporary edge connection is established between the nodes at that time series position, with the edge weight being the corresponding element value. The preset association threshold is a pre-defined value used to determine whether an association exists between two nodes.

[0074] As one implementation method, step S330, determining the connection relationship between nodes at each time series position based on the correlation strength matrix, can be specifically implemented as the following steps S331~S335:

[0075] Step S331: Perform matrix standardization on the correlation strength matrix, adjusting the matrix element values ​​to a preset range to make the correlation strength matrices at different time series positions comparable.

[0076] The purpose of matrix standardization of the association strength matrix is ​​to adjust the element values ​​of the association strength matrix at different time series locations to a common numerical range, making the association strength matrices at different time points comparable. Matrix standardization can be performed using various methods, such as the min-max standardization method, which adjusts the matrix element values ​​to the [0,1] interval. In implementation, the minimum and maximum values ​​in the association strength matrix can be found first. Then, the minimum value is subtracted from each element in the matrix, and the result is divided by the difference between the maximum and minimum values ​​to obtain the standardized matrix element values.

[0077] Step S332: The standardized correlation strength matrix is ​​processed using the nonmaximum suppression algorithm. A preset number of elements with the largest values ​​in each row are retained, and the other elements are set to zero to obtain a sparse correlation matrix.

[0078] Non-maximum suppression (NMS) is used to process the standardized association strength matrix to reduce redundant information and improve computational efficiency. NMS retains a preset number of elements with the largest values ​​in each row, setting all other elements to zero, resulting in a sparse association matrix. The preset number can be set according to specific needs, for example, 3 or 5. In implementation, each row of the standardized association strength matrix is ​​traversed, and the preset number of elements with the largest values ​​in each row are found, with all other elements set to zero.

[0079] Step S333: Perform symmetry processing on the sparse correlation matrix. If the element in the i-th row and j-th column of the matrix is ​​non-zero and the element in the j-th row and i-th column is zero, then assign the value of the element in the j-th row and i-th column to the element in the i-th row and j-th column.

[0080] The purpose of symmetrizing the sparse incidence matrix is ​​to ensure that the connections between nodes are symmetric. In a time-varying dependency network, if there is a connection between node A and node B, then there should also be a connection between node B and node A. Therefore, if the element in the i-th row and j-th column of the matrix is ​​non-zero and the element in the j-th row and i-th column is zero, then the element in the j-th row and i-th column is assigned the value of the element in the i-th row and j-th column.

[0081] In the implementation process, the sparse correlation matrix can be traversed to check the symmetry of the matrix. If an asymmetric element is found, an assignment operation is performed.

[0082] Step S334: Traverse the non-zero elements in the symmetricized sparse correlation matrix, and record the row index and column index corresponding to each non-zero element. The row index and column index correspond to the identifiers of the two nodes, respectively.

[0083] The non-zero elements in the symmetricized sparse incidence matrix are traversed, and the row and column indices corresponding to each non-zero element are recorded. The purpose is to determine the connection relationships between nodes. The row and column indices correspond to the identifiers of two nodes, respectively. By recording these indices, the connection information between nodes can be obtained.

[0084] In the implementation process, all non-zero elements can be found by traversing the symmetricized sparse correlation matrix, and their row and column indices can be recorded.

[0085] Step S335: Use the node identifier pair corresponding to each non-zero element as the connection relationship between nodes, and the non-zero element value as the edge weight of the corresponding connection relationship to obtain the connection relationship between nodes at this time series position.

[0086] By using the node identifier pairs corresponding to each non-zero element as the connection relationships between nodes, and the non-zero element values ​​as the edge weights of the corresponding connections, the connection relationships between nodes at that time series position can be obtained. In this way, the sparse correlation matrix can be transformed into connection relationships between nodes, providing a foundation for constructing time-varying dependency networks.

[0087] Step S340: Integrate the temporary edge connections and edge weights of all time series positions in time series order to obtain a time-varying edge set. The time-varying edge set contains the edge connection states between nodes at different time series positions and the corresponding edge weight values.

[0088] By integrating the temporary edge connections and weights at all time series positions in chronological order, a time-varying edge set is obtained. This set contains the edge connection states and corresponding weight values ​​between nodes at different time series positions, reflecting the temporal changes in the connection relationships between nodes. By integrating the edge connection information at all time points, a complete edge structure of the time-varying dependency network can be constructed.

[0089] In the implementation process, the connections between nodes at each time point can be stored in a list, with each element in the list corresponding to a connection at a time point. Then, these lists are arranged in chronological order to obtain a time-varying edge set.

[0090] Step S350: Extract the feature vector corresponding to each node at each time series position. The node feature vector is obtained by performing element-wise multiplication of the first dimension vector of the node in the association strength tensor with the feature vector of the corresponding time series position of the node's coupling feature sequence.

[0091] The goal of extracting the feature vectors of each node at each time series position is to obtain the feature information of each node at different time points. The node feature vector is obtained by element-wise multiplying the first dimension vector of the node in the association strength tensor with the feature vector of the node's coupling feature sequence at the corresponding time series position. This method can fuse the node's association information with its own feature information, resulting in more comprehensive node features.

[0092] As one implementation method, step S350, extracting the feature vector corresponding to each node at each time series position, can be specifically implemented as the following steps S351~S356:

[0093] Step S351: Extract the first dimension vector corresponding to the target node from the association strength tensor to obtain the node association feature vector. The dimension of the node association feature vector is consistent with the third dimension of the association strength tensor, and the element value is the degree of association between the target node and other nodes.

[0094] The first-dimensional vector corresponding to the target node is extracted from the association strength tensor to obtain the node association feature vector. The dimension of the node association feature vector is consistent with the third dimension of the association strength tensor, and its element values ​​represent the degree of association between the target node and other nodes. By extracting the node association feature vector, the association relationship between the target node and other nodes can be understood.

[0095] In the implementation process, the first dimension vector can be extracted based on the position of the target node in the association strength tensor to obtain the node association feature vector.

[0096] Step S352: Extract the coupling feature sub-vectors of the target node at the corresponding time series position from the coupling feature sequence. The coupling feature sub-vectors contain the carbon emission monitoring features and related impact features of the target node.

[0097] The coupling feature sub-vectors of the target node at the corresponding time series position are extracted from the coupling feature sequence. These coupling feature sub-vectors contain the carbon emission monitoring characteristics and associated impact characteristics of the target node. These features can reflect the actual state of the target node at that time point, such as carbon emissions, outdoor temperature, and indoor population density.

[0098] During implementation, the corresponding coupled feature sub-vectors can be extracted from the coupled feature sequence based on the target node and time series position.

[0099] Step S353: Perform dimension adaptation processing on the node association feature vector and coupling feature sub-vector. Adjust the dimensions of the two vectors to the same dimension through linear transformation to obtain the adapted association vector and the adapted coupling vector.

[0100] Dimensional adaptation is performed on the node association feature vector and the coupling feature sub-vector to ensure that the two vectors have the same dimension for element-wise multiplication. A linear transformation adjusts the dimensions of the two vectors to be the same, resulting in an adapted association vector and an adapted coupling vector. This linear transformation can be achieved through matrix multiplication. For example, assuming the node association feature vector is A, the coupling feature sub-vector is B, and the dimension adaptation matrix is ​​W, then the adapted association vector A' = A × W, and the adapted coupling vector B' = B × W.

[0101] Step S354: Perform element-wise multiplication on the adaptation association vector and the adaptation coupling vector to obtain the preliminary node feature vector.

[0102] Element-wise multiplication of the adaptation association vector and the adaptation coupling vector yields a preliminary node feature vector. Element-wise multiplication involves multiplying corresponding elements of two vectors to obtain a new vector. In this way, the association information of a node can be fused with its own feature information to obtain more comprehensive node features.

[0103] In the implementation process, the preliminary node feature vector can be obtained by traversing each element of the adaptation association vector and the adaptation coupling vector and multiplying the corresponding elements.

[0104] Step S355: Perform feature smoothing on the initial node feature vectors. High-frequency noise in the vectors is eliminated by moving average filtering. The window size of the moving average filtering is the preset feature smoothing window length.

[0105] The purpose of smoothing the initial node feature vectors is to eliminate high-frequency noise and improve their stability. This smoothing can be achieved through moving average filtering, with the window size set to a preset feature smoothing window length. During the moving average filtering process, the average value of each element within the window is calculated, and this average value is used as the new value for the center element of the window. By continuously sliding the window, the entire vector can be smoothed. Specifically, the initial node feature vectors can be traversed, and the average value of each element within each window can be calculated according to the preset window size, with the average value used as the new value for the center element of the window.

[0106] Step S356: Use the smoothed preliminary node feature vector as the node feature vector of the target node at this time series position.

[0107] Using the smoothed initial node feature vector as the node feature vector of the target node at that time series position, the smoothed node feature vector is more stable and can more accurately reflect the feature information of the node at that time point.

[0108] Step S360: Arrange the node feature vectors in time sequence to obtain the node feature vector time series.

[0109] Arranging node feature vectors in chronological order yields the node feature vector time series. This time series reflects the changes in the feature information of each node at different time points and is a crucial representation of node features in time-varying networks. By arranging the node feature vectors in chronological order, a complete node feature sequence can be constructed, providing a foundation for subsequent analysis. In implementation, the node feature vectors of each node at different time points can be stored in a list, where each element corresponds to a node feature vector at a given time point. Then, these lists are arranged in chronological order to obtain the node feature vector time series.

[0110] Step S370: Calculate the rate of change of the edge weight of each edge in the time-varying edge set at adjacent time series positions to obtain the edge weight evolution gradient. The evolution gradient is the difference between the edge weight at the current time series position and the edge weight at the previous time series position.

[0111] The edge weight evolution gradient is obtained by calculating the rate of change of the edge weight at adjacent time series positions in the time-varying edge set. The edge weight evolution gradient reflects the change of edge weights over time, capturing the dynamic changes in the connections between nodes. The evolution gradient is the difference between the edge weight at the current time series position and the edge weight at the previous time series position. Specifically, the time-varying edge set can be traversed, and for each edge, its edge weight values ​​at adjacent time series positions can be obtained, and their differences can be calculated to obtain the edge weight evolution gradient. The edge weight evolution gradient of each edge can be stored in a list, where each element corresponds to the evolution gradient of one edge. For easier subsequent processing, these evolution gradients can also be associated with the corresponding edge information (such as the starting node, ending node, etc.) for storage.

[0112] Step S380: Integrate the time series of node feature vectors and the evolution gradient of edge weights to obtain time-varying network parameters.

[0113] Integrating the time series of node feature vectors and the evolutionary gradients of edge weights yields complete parameters for time-dependent networks. These parameters comprehensively describe the changes in the characteristics and structure of the time-dependent network over time. The time series of node feature vectors reflects the characteristic information of each network node at different time points, demonstrating the dynamic changes of the nodes themselves; while the evolutionary gradients of edge weights reflect the rate of change of the connections between nodes over time, demonstrating the dynamic characteristics of the network structure. Integrating these two parameters provides a rich and accurate information foundation for the analysis and prediction of building carbon emission data.

[0114] During the integration process, the node feature vector time series and edge weight evolution gradients can be stored in a data structure, such as a dictionary. The dictionary keys can be set to "node feature vector time series" and "edge weight evolution gradients," respectively, with the corresponding values ​​being the previously obtained lists of node feature vector time series and edge weight evolution gradients. This data structure facilitates subsequent access and manipulation of time-varying network parameters.

[0115] Step S400: Track the multi-scale evolution trajectory by time-varying network parameters to generate a multi-scale evolution trajectory set. The multi-scale evolution trajectory set includes the evolution trajectory of node feature vectors and the gradient change trajectory of time-varying edge weights between nodes at short-term, medium-term and long-term time scales.

[0116] Tracking multi-scale evolution trajectories allows for in-depth analysis of changes in factors related to building carbon emissions across different time scales. Short-term time scales can capture rapid fluctuations and immediate changes in carbon emissions, medium-term time scales help identify cyclical changes and trends in carbon emissions, and long-term time scales can reveal macroscopic patterns and overall development trends in carbon emissions.

[0117] As one implementation method, step S400, which involves tracking multi-scale evolution trajectories through time-varying network parameters to generate a set of multi-scale evolution trajectories, can be specifically implemented as follows: S410~S480:

[0118] Step S410: Determine the multi-scale time window set, which includes short-term time windows, medium-term time windows, and long-term time windows. The window length of each time window is a different proportion of the number of positions in the continuous time series.

[0119] Determining a set of multi-scale time windows is fundamental to tracking multi-scale evolutionary trajectories. Different time window lengths can be set according to actual needs and data characteristics. Short-term time windows are relatively short, used to capture short-term changes; medium-term time windows are of moderate length, used to analyze periodic and phased changes; and long-term time windows are longer, used to observe long-term trends and patterns. For example, the short-term time window can be set to 10% of the number of positions in a continuous time series, the medium-term time window to 30%, and the long-term time window to 60%. Such settings can be flexibly adjusted according to different building types, the characteristics of carbon emission data, and the purpose of analysis.

[0120] Step S420: Based on the time series of node feature vectors in the time-varying network parameters, divide the node feature vector time series into windows according to the multi-scale time window set to obtain the short-term feature window sequence, medium-term feature window sequence and long-term feature window sequence of each node. The feature window sequence is composed of node feature vectors at multiple consecutive time series positions.

[0121] Dividing the node feature vector time series into windows using a multi-scale time window set allows for the segmentation of time series data into time periods of different scales, facilitating the analysis of feature changes at each scale. Each feature window sequence contains node feature vectors at multiple consecutive time series positions, reflecting the feature changes of nodes within the corresponding time window. Analyzing feature window sequences at different scales provides a more comprehensive understanding of the dynamic changes in node features.

[0122] As one implementation method, step S420 involves dividing the node feature vector time series based on the node feature vector time series in the time-varying network parameters into windows according to a multi-scale time window set, to obtain the short-term feature window sequence, medium-term feature window sequence, and long-term feature window sequence for each node. Specifically, this can be implemented as the following steps S421~S426:

[0123] Step S421: Obtain the total time length of the node feature vector time series, where the total time length is the total number of time series positions.

[0124] Obtaining the total time length of the node feature vector time series is a prerequisite for window segmentation. By counting the total number of time series positions, the range of the entire time series can be determined, providing a basis for subsequent calculations of the length and sliding step of different time windows. The total time length can be obtained by traversing the list of node feature vector time series and counting the number of elements in the list.

[0125] Step S422: Determine the sliding step size of the short-term time window based on the total time length and the window length of the short-term time window. The sliding step size is the number of time series positions the window moves, so that there is a preset proportion of overlapping area between adjacent short-term time windows.

[0126] The purpose of determining the sliding step size for the short-term time window is to enable window movement and overlap during window division. A preset overlap ratio ensures data continuity and information integrity, preventing information loss or abrupt changes. Determining the sliding step size requires considering both the length of the short-term time window and the preset overlap ratio. For example, if the short-term time window is 10 time series positions long and the preset overlap ratio is 50%, then the sliding step size can be set to 5 time series positions. In this way, adjacent short-term time windows will have an overlap of 5 time series positions.

[0127] Step S423: Starting from the beginning position of the node feature vector time series, extract subsequences according to the short-term time window length and sliding step size to obtain multiple short-term feature windows. Arrange these windows in the extraction order to obtain a short-term feature window sequence.

[0128] Starting from the beginning of the node feature vector time series, subsequences are extracted according to a determined short-term time window length and sliding step size. Each time, a subsequence of the short-term time window length is extracted, and then the window is moved forward by the sliding step size, continuing this process until the entire time series has been traversed. The extracted short-term feature windows are then arranged in the order of extraction to obtain the short-term feature window sequence.

[0129] Step S424: Determine the intermediate sliding step size based on the window length of the intermediate time window and the preset overlap ratio, and obtain the intermediate feature window sequence using the same truncation method.

[0130] The method for determining the intermediate sliding step size is similar to that for the short-term sliding step size, calculated based on the window length of the intermediate time window and a preset overlap ratio. Then, using the same method as for extracting the short-term feature window, intermediate feature windows are extracted from the node feature vector time series and arranged sequentially to obtain the intermediate feature window sequence. The length and overlap ratio of the intermediate time window can be adjusted according to actual needs to suit different analytical purposes.

[0131] Step S425: Determine the long-term sliding step size based on the window length of the long-term time window and the preset overlap ratio, and obtain the long-term feature window sequence using the same truncation method.

[0132] Similarly, the long-term sliding step size is determined based on the window length of the long-term time window and the preset overlap ratio. Then, long-term feature windows are extracted from the node feature vector time series using the same truncation method. These windows are arranged in order to obtain the long-term feature window sequence. The length of the long-term time window is usually relatively long, and the overlap ratio can also be adjusted as needed to accurately capture long-term trends.

[0133] Step S426: Perform boundary processing on the windows in each feature window sequence. If the length of the last window is less than the length of the corresponding time window, then make up the window length by copying the feature vector of the last node so that all windows have the same length.

[0134] Boundary processing is performed on windows in each feature window sequence to ensure that all windows have the same length, facilitating subsequent analysis and processing. If the length of the last window is shorter than the length of the corresponding time window, the window length is padded by copying the feature vector of the last node. This ensures that each window contains the same number of node feature vectors, avoiding analysis errors caused by inconsistent window lengths.

[0135] Step S430: Perform trajectory fitting on the short-term feature window sequence of each node, and obtain the short-term feature evolution trajectory through a polynomial curve fitting algorithm. The short-term feature evolution trajectory represents the changing trend of the node feature vector within the short-term time window.

[0136] Trajectory fitting of short-term feature window sequences can reveal the changing trends of nodal feature vectors within a short time window. Polynomial curve fitting is a commonly used method, which approximates the changes in nodal feature vectors over time using polynomial functions. The short-term feature evolution trajectories obtained through fitting can visually demonstrate the rise, fall, or fluctuation of nodal features in the short term, providing a basis for predicting short-term carbon emission changes.

[0137] As one implementation method, step S430 involves trajectory fitting of the short-term feature window sequence for each node, obtaining the short-term feature evolution trajectory through a polynomial curve fitting algorithm. Specifically, this can be implemented as follows: steps S431~S437:

[0138] Step S431: Select a short-term feature window from the short-term feature window sequence as the target window, extract the node feature vectors of all time series positions within the target window, and obtain the window feature vector set.

[0139] The purpose of selecting a short-term feature window from the short-term feature window sequence as the target window is to perform detailed analysis of the node feature vectors within that window. The node feature vectors for all time series positions within the target window are extracted and stored in a set, resulting in the window feature vector set. This set contains detailed information about the node features within the target window and serves as the foundational data for trajectory fitting.

[0140] Step S432: Convert each node feature vector in the window feature vector set into a feature value sequence. The feature value sequence is a one-dimensional sequence obtained by arranging the feature values ​​of each dimension of the vector in dimensional order.

[0141] The purpose of converting each node feature vector in the window feature vector set into a sequence of feature values ​​is to transform multidimensional node feature vectors into a one-dimensional sequence, facilitating polynomial curve fitting. By arranging the feature values ​​of each dimension of the vector in dimensional order, a one-dimensional feature value sequence is obtained. This transformation simplifies complex multidimensional features into a one-dimensional numerical sequence, improving the efficiency and accuracy of fitting.

[0142] Step S433: Perform time index mapping on the feature value sequence, mapping the time series position corresponding to each feature value to a relative time index within the window, with the relative time index incrementing from 0 at the beginning of the window.

[0143] The purpose of mapping the feature value sequence to a time index is to associate the feature values ​​with their relative time positions within a window. Each feature value's corresponding time sequence position is mapped to a relative time index within the window, with the relative time index incrementing from 0 at the beginning of the window. This establishes a clear correspondence between the feature value sequence and time, facilitating subsequent fitting using polynomial functions.

[0144] Step S434: Use a polynomial function of a preset order to perform curve fitting on the feature value sequence and the corresponding relative time index to obtain the fitting polynomial for each dimension of feature value. The independent variable of the fitting polynomial is the relative time index, and the dependent variable is the feature value.

[0145] A polynomial function of a preset order is used to perform curve fitting on the feature value sequence and its corresponding relative time index. The aim is to find a suitable polynomial function to approximate the change of the node feature vector over time. The preset order can be selected according to the actual situation. The higher the order, the higher the complexity of the polynomial function, and the higher the fitting accuracy may be, but overfitting may also occur. The fitted polynomial for each dimension of the feature value obtained by fitting has the relative time index as the independent variable and the feature value as the dependent variable. This polynomial can be used to predict the change of node features within a window.

[0146] Step S435: Obtain the fitted feature value sequence by calculating the predicted feature values ​​corresponding to all relative time indices within the window using a fitted polynomial.

[0147] The purpose of this study is to verify the fitting effect by calculating the predicted feature values ​​corresponding to all relative time indices within a window using a fitting polynomial. Substituting the relative time indices within the window into the fitting polynomial yields the corresponding predicted feature values. These predicted feature values ​​are then arranged in order of their relative time indices to obtain the fitted feature value sequence. This sequence can be compared with the original feature value sequence to evaluate the accuracy of the fit.

[0148] Step S436: Calculate the fitting error between the fitted feature value sequence and the original feature value sequence. If the fitting error is greater than the preset error threshold, increase the polynomial order and refit until the fitting error is less than the preset error threshold.

[0149] The purpose of calculating the fitting error between the fitted feature value sequence and the original feature value sequence is to evaluate the quality of the fit. The fitting error can be obtained by calculating the absolute value or sum of squares of the differences between the two sequences. If the fitting error is greater than a preset error threshold, it indicates that the current fitting effect is not ideal, and the polynomial order needs to be increased for refitting. By continuously adjusting the polynomial order until the fitting error is less than the preset error threshold, the accuracy of the fit is ensured.

[0150] Step S437: Use the final fitted polynomial as a short-term feature evolution trajectory segment of the target window, arrange all short-term feature evolution trajectory segments in window order, and obtain the short-term feature evolution trajectory.

[0151] By using the final fitted polynomial as a segment of the short-term feature evolution trajectory for the target window, and arranging all short-term feature evolution trajectory segments in window order, the short-term feature evolution trajectory can be obtained. This trajectory can comprehensively show the changing trend of nodal feature vectors within a short-term time window, providing strong support for analyzing short-term carbon emission changes.

[0152] Step S440: Use the same method to perform trajectory fitting on the intermediate feature window sequence and the long-term feature window sequence to obtain the intermediate feature evolution trajectory and the long-term feature evolution trajectory.

[0153] The same method used for processing short-term feature window sequences was employed to fit trajectories to the medium-term and long-term feature window sequences. A polynomial curve fitting algorithm was used to obtain the medium-term and long-term feature evolution trajectories, respectively. The medium-term feature evolution trajectory reflects the changing trend of node feature vectors over a medium-term time scale, while the long-term feature evolution trajectory reveals the development pattern of node features over a long-term time scale. These two trajectories, together with the short-term feature evolution trajectory, constitute a multi-scale feature evolution trajectory.

[0154] Step S450: Extract the gradient values ​​of the edge weights between nodes at each time series position from the edge weight evolution gradient in the time-varying network parameters, and divide them into short-term gradient window sequences, medium-term gradient window sequences and long-term gradient window sequences according to the multi-scale time window set.

[0155] The gradient values ​​of the edge weights between nodes at various time series positions are extracted from the edge weight evolution gradients in the time-varying network parameters. These gradient values ​​reflect the rate of change of the inter-node connections over time. These gradient values ​​are then divided into short-term, medium-term, and long-term gradient window sequences according to a multi-scale time window set. Similar to the windowing of node feature vectors, gradient window sequences at different scales can capture the changes in edge weights in the short, medium, and long term, respectively, providing a foundation for analyzing the dynamic changes in inter-node connections.

[0156] Step S460: Perform gradient accumulation processing on the short-term gradient window sequence of each node edge, calculate the sum of gradient values ​​within the window, and obtain the short-term gradient change trajectory. The short-term gradient change trajectory represents the cumulative change of edge weight within the short-term time window.

[0157] By accumulating the gradients within a short-term gradient window sequence for each node's edges, and summing the gradient values ​​within the window, the short-term gradient trajectory can be obtained. This trajectory represents the cumulative change in edge weights within a short-term time window, reflecting the dynamic changes in the connectivity between nodes in the short term. The short-term gradient trajectory can help analysts quickly understand the increasing or decreasing trends of edge weights in the short term, providing a basis for predicting short-term carbon emission correlation changes.

[0158] Step S470: Use the same method to perform gradient accumulation processing on the intermediate gradient window sequence and the long-term gradient window sequence to obtain the intermediate gradient change trajectory and the long-term gradient change trajectory.

[0159] The same method used for processing short-term gradient window sequences is applied to accumulate gradients in medium-term and long-term gradient window sequences. The sum of gradient values ​​within the medium-term and long-term windows is calculated separately to obtain the medium-term and long-term gradient change trajectories. The medium-term gradient change trajectory reflects the cumulative changes of edge weights over the medium-term time scale, while the long-term gradient change trajectory reveals the development trend of edge weights over the long-term time scale.

[0160] Step S480: Integrate the short-term feature evolution trajectory, medium-term feature evolution trajectory, and long-term feature evolution trajectory of all nodes, as well as the short-term gradient change trajectory, medium-term gradient change trajectory, and long-term gradient change trajectory of all edges, to obtain a multi-scale evolution trajectory set.

[0161] By integrating the short-term, medium-term, and long-term feature evolution trajectories of all nodes, as well as the short-term, medium-term, and long-term gradient change trajectories of all edges, a comprehensive set of multi-scale evolution trajectories can be obtained. This set contains the feature evolution trajectories of network nodes at different time scales and the gradient change trajectories of time-varying edge weights between nodes, providing rich and comprehensive information for in-depth analysis of multi-scale changes in building carbon emission data. These trajectories can be stored in a data structure, such as a list or dictionary, for convenient subsequent access and use.

[0162] Step S500: Perform critical path identification using the attention mechanism through the multi-scale evolution trajectory set, and output the carbon emission data processing results. The critical path is the sequence of inter-node association paths that dominate the contribution of the overall carbon emission characteristic evolution in the multi-scale evolution trajectory through attention weight.

[0163] Critical path identification using attention mechanisms can identify the dominant inter-node path sequences that contribute significantly to the overall evolution of carbon emission characteristics from a multi-scale evolutionary trajectory set. Attention mechanisms can assign attention weights based on the importance of different trajectories, and by selecting paths with larger attention weights, critical paths are identified. These critical paths reflect the most critical relationships and change paths in the building carbon emission system.

[0164] As one implementation method, step S500 involves performing critical path identification using an attention mechanism through a multi-scale evolutionary trajectory set, and outputting carbon emission data processing results. Specifically, this can be implemented as the following steps S510~S570:

[0165] Step S510: Construct a multi-scale attention network. The multi-scale attention network includes a trajectory encoding layer, a cross-scale attention layer, and a path scoring layer. The trajectory encoding layer is used to encode the features of the multi-scale evolution trajectory. The cross-scale attention layer is used to calculate the attention weights between trajectories of different scales. The path scoring layer is used to score the contribution of the associated paths between nodes.

[0166] Constructing a multi-scale attention network is a core step in critical path identification. The multi-scale attention network consists of a trajectory encoding layer, a cross-scale attention layer, and a path scoring layer. The trajectory encoding layer encodes the features of multi-scale evolving trajectories, converting trajectory data at different scales into feature representations suitable for network processing. The cross-scale attention layer, by calculating attention weights between trajectories at different scales, captures the correlation and importance between trajectories at different scales. The path scoring layer scores the contribution of inter-node paths, evaluating the contribution of each path to the overall carbon emission characteristic evolution based on the features of each node and edge in the path, as well as the attention weights.

[0167] The trajectory encoding layer can be implemented using a convolutional neural network (CNN), which can effectively extract spatial and temporal features from trajectory data. The cross-scale attention layer can employ a self-attention mechanism, assigning attention weights by calculating the similarity between trajectories at different scales. The path scoring layer can use a graph attention mechanism, scoring the associated paths between nodes based on the features of nodes and edges, as well as the attention weights.

[0168] Step S520: Input the multi-scale evolution trajectory set into the trajectory encoding layer of the multi-scale attention network, and extract features from the feature evolution trajectory and gradient change trajectory at each scale through the convolutional neural network to obtain multi-scale trajectory encoding features.

[0169] A set of multi-scale evolutionary trajectories is input into the trajectory encoding layer of a multi-scale attention network. A convolutional neural network (CNN) is then used to extract features from the feature evolution trajectories and gradient change trajectories at each scale. Through operations such as convolution and pooling, the CNN can extract representative features from the trajectory data. The resulting multi-scale trajectory encoding features contain key information about trajectories at different scales, providing a foundation for cross-scale attention computation and path scoring.

[0170] In the implementation process, a set of multi-scale evolutionary trajectories is input into a convolutional neural network according to a certain format. Appropriate parameters such as kernel size, stride, and number of convolutional layers are set for feature extraction. After processing by the convolutional neural network, multi-scale trajectory encoding features are obtained.

[0171] Step S530: Input the multi-scale trajectory encoding features into the cross-scale attention layer, calculate the correlation weights between the short-term trajectory encoding features, the medium-term trajectory encoding features and the long-term trajectory encoding features through the self-attention mechanism, and generate cross-scale fusion features. The cross-scale fusion features integrate the key information of trajectories at different scales.

[0172] Multi-scale trajectory encoding features are input into a cross-scale attention layer, and a self-attention mechanism is used to calculate the association weights between short-term, medium-term, and long-term trajectory encoding features. The self-attention mechanism assigns corresponding attention weights to each feature by calculating the similarity between trajectory encoding features at different scales. Based on these attention weights, the trajectory encoding features at different scales are weighted and combined to generate cross-scale fused features. These cross-scale fused features integrate key information from trajectories at different scales, providing a more comprehensive reflection of the multi-scale changes in building carbon emissions.

[0173] In the implementation of the self-attention mechanism, the multi-scale trajectory encoding features are first linearly transformed to obtain the query vector, key vector, and value vector. Then, the similarity between the query vector and the key vector is calculated, and the similarity is converted into attention weights using the softmax function. Finally, the value vectors are weighted and summed according to the attention weights to obtain the cross-scale fused features.

[0174] Step S540: Extract all possible inter-node association paths from the time-varying dependency network. An association path is a sequence of nodes formed by connecting multiple nodes through edges.

[0175] The purpose of extracting all possible inter-node association paths from a time-varying dependency network is to evaluate and filter all possible paths. An association path is a sequence of nodes connected by edges, reflecting the relationships and information transfer paths between nodes. All possible inter-node association paths can be found from the time-varying dependency network using graph traversal algorithms such as Depth-First Search (DFS) or Breadth-First Search (BFS). These paths are then stored in a list, providing a basis for path scoring.

[0176] Step S550: Input the cross-scale fusion features and the inter-node association paths into the path scoring layer, and calculate the contribution score of each association path through the graph attention mechanism. The contribution score is obtained by weighting the importance of the trajectory features of each node in the path and the importance of the gradient changes of each edge.

[0177] The cross-scale fusion features and inter-node association paths are input into the path scoring layer, and a graph attention mechanism is used to calculate the contribution score of each association path. The graph attention mechanism evaluates each path based on the trajectory features of each node and the gradient change features of each edge, combined with the attention weights in the cross-scale fusion features. The contribution score is weighted based on the importance of the trajectory features of each node and the importance of the gradient changes of each edge, reflecting the degree to which each path contributes to the overall evolution of carbon emission characteristics.

[0178] As one implementation method, step S550 involves inputting the cross-scale fusion features and the inter-node association paths into the path scoring layer, and calculating the contribution score of each association path through a graph attention mechanism. Specifically, this can be implemented as follows: steps S551 to S557.

[0179] Step S551: Decompose the inter-node association path into a path node sequence and a path edge sequence. The path node sequence is the sequence of nodes contained in the path arranged in order, and the path edge sequence is the sequence of edges contained in the path arranged in order.

[0180] The inter-node paths are decomposed into a path node sequence and a path edge sequence to process the features of nodes and edges in the path separately. The path node sequence contains information about all nodes in the path, arranged in order; the path edge sequence contains information about all edges in the path, also arranged in order. This decomposition makes it easier to extract the features of nodes and edges, providing a foundation for subsequent scoring calculations.

[0181] Step S552: Extract the node trajectory features corresponding to each node in the path node sequence from the cross-scale fusion features to obtain the node feature sequence.

[0182] The node trajectory features corresponding to each node in the path node sequence are extracted from the cross-scale fusion features. These features are then arranged in node order to obtain the node feature sequence. The node trajectory features reflect the feature evolution of nodes at different scales and are an important basis for evaluating the contribution of nodes to the path.

[0183] Step S553: ​​Extract the edge gradient features corresponding to each edge in the path edge sequence to obtain the edge feature sequence.

[0184] The edge gradient features corresponding to each edge in the path edge sequence are extracted, and these features are arranged in the order of the edges to obtain the edge feature sequence. The edge gradient features reflect the gradient changes of the edge weights at different scales and are an important basis for evaluating the contribution of edges to the path.

[0185] Step S554: Input the node feature sequence and edge feature sequence into the node attention module and edge attention module of the graph attention mechanism, and calculate the node attention weight of each node and the edge attention weight of each edge. The node attention weight represents the importance of the node in the path, and the edge attention weight represents the importance of the edge in the path.

[0186] The node feature sequences and edge feature sequences are input into the node attention module and edge attention module of the graph attention mechanism, respectively. The node attention module calculates the node attention weight for each node based on the node feature sequence, and the edge attention module calculates the edge attention weight for each edge based on the edge feature sequence. The node attention weights and edge attention weights characterize the importance of nodes and edges in the path, respectively, allowing for a more accurate assessment of the path's contribution. The node attention module and edge attention module can be implemented using linear transformations and a softmax function. A linear transformation converts the node and edge features into attention scores, and then a softmax function converts these scores into attention weights.

[0187] Step S555: The node feature vectors in the node feature sequence are weighted and summed according to the node attention weights to obtain the total features of the path nodes.

[0188] The node feature vectors in the node feature sequence are weighted and summed according to the node attention weights to obtain the total feature of the path nodes. By weighting and summing, the node feature vectors and node attention weights can be combined to obtain a comprehensive node feature representation that reflects the overall characteristics and importance of all nodes in the path.

[0189] Step S556: The edge feature vectors in the edge feature sequence are weighted and summed according to the edge attention weights to obtain the total feature of the path edges.

[0190] The edge feature vectors in the edge feature sequence are weighted and summed according to the edge attention weights to obtain the total edge features of the path. Similarly, by weighted summation, the edge feature vectors and edge attention weights can be combined to obtain a comprehensive edge feature representation that reflects the overall characteristics and importance of all edges in the path.

[0191] Step S557: Concatenate the total features of the path nodes and the total features of the path edges, and map them to scalar values ​​through a fully connected layer to obtain the contribution score of the associated path.

[0192] The total features of path nodes and total features of path edges are concatenated, merging the two feature vectors into a larger feature vector. Then, a fully connected layer maps the concatenated feature vector into a scalar value, yielding a contribution score for the associated path. The fully connected layer transforms a high-dimensional feature vector into a scalar value, reflecting the degree to which the associated path contributes to the overall evolution of carbon emission characteristics.

[0193] Step S560: Sort the contribution scores of all associated paths in descending order, and select the pre-defined number of associated paths that rank highest as critical paths.

[0194] The contribution scores of all associated paths are sorted in descending order, arranging the paths from highest to lowest contribution. A predetermined number of associated paths with the highest contribution are selected as critical paths; the predetermined number can be set according to actual needs and analysis objectives. These critical paths are those that contribute the most to the overall evolution of carbon emission characteristics.

[0195] Step S570: Arrange the critical paths in order of contribution score to generate carbon emission data processing results.

[0196] The critical paths are arranged in order of contribution score to generate carbon emission data processing results. These results contain key information obtained after processing building carbon emission data, namely, the sequence of inter-node association paths that dominate the evolution of overall carbon emission characteristics. By analyzing these critical paths, targeted strategies and suggestions can be provided for the management and control of building carbon emissions, such as optimizing the operational status of nodes and edges on the critical paths to reduce carbon emissions.

[0197] All algorithms mentioned in the embodiments of this invention can be found in relevant existing technologies. To save space, they will not be elaborated upon in the embodiments of this invention. Furthermore, those skilled in the art can supplement the details based on common knowledge in the field when implementing the solutions of this invention. For example, they can use normalization to eliminate dimensional conflicts before feature fusion, use interpolation to eliminate dimensional differences, reasonably set thresholds based on historical data, experience, or business scenario requirements, train the model based on a general model training method, set the number of layers in the model structure based on actual needs, select activation functions, etc. This invention will not provide redundant descriptions of overly detailed implementation processes here.

[0198] Figure 2 A hardware entity diagram of a computer system provided as an embodiment of the present invention, such as... Figure 2 As shown, the hardware entity of the computer system 1000 includes a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can run on the processor 1001, and the processor 1001 executes the program to implement the steps in the method of any of the above embodiments.

[0199] The memory 1002 stores computer programs that can run on the processor. The memory 1002 is configured to store instructions and applications that can be executed by the processor 1001. It can also cache data to be processed or already processed (e.g., image data, audio data, voice communication data, and video communication data) of the processor 1001 and various modules in the computer system 1000. It can be implemented by flash memory or random access memory (RAM).

[0200] When processor 1001 executes a program, it implements the steps of any of the above-mentioned deep learning-based building carbon emission data processing methods. Processor 1001 typically controls the overall operation of computer system 1000.

[0201] The above description is merely an embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based method for processing building carbon emission data, characterized in that, include: Acquire a building carbon emission data set, which includes a carbon emission monitoring sequence and related impact sequence of a building functional system cluster in a continuous time series. The building functional system cluster is a combination of functional units that generate carbon emissions during building operation, and the related impact sequence is a sequence of internal and external environmental parameters that affect the carbon emission characteristics of the functional units as a time series. Feature association mining is performed on the building carbon emission dataset to generate an association strength tensor. The first dimension of the association strength tensor corresponds to the functional unit, the second dimension corresponds to the time series position, and the third dimension corresponds to the feature association type. The tensor elements represent the degree of association between the carbon emission characteristics of the corresponding functional unit at the corresponding time series position and the carbon emission characteristics of other functional units. A time-varying dependency network is constructed based on the correlation strength tensor, and the parameters of the time-varying dependency network are obtained. By tracking multi-scale evolution trajectories through the time-varying network parameters, a set of multi-scale evolution trajectories is generated; The attention mechanism is used to identify critical paths through the multi-scale evolution trajectory set, and the carbon emission data processing results are output. The step of performing feature association mining on the building carbon emission dataset to generate an association strength tensor includes: The carbon emission monitoring subsequences and associated impact subsequences of each functional unit are separated from the building carbon emission data set to obtain a functional unit data sequence set. Each element in the functional unit data sequence set is a combination sequence of the carbon emission monitoring subsequence and associated impact subsequence corresponding to a functional unit. The combined sequences in the set of functional unit data sequences are subjected to feature coupling processing to generate coupled feature sequences. The feature coupling processing interweaves the carbon emission monitoring subsequence and the associated impact subsequence according to the feature dimension of the time series position, so that the feature vector of each time series position simultaneously contains carbon emission monitoring features and associated impact features. A correlation mining network is constructed, which includes a sequence encoding layer, a cross-correlation layer, and a tensor output layer. The sequence encoding layer is used to extract time series features from coupled feature sequences. The cross-correlation layer is used to calculate the correlation between coupled feature sequences of different functional units. The tensor output layer is used to integrate the correlation results to generate a tensor. The set of data sequences of the functional units is input into the sequence encoding layer of the association mining network. The coupled feature sequences of each functional unit are encoded by stacked gated recurrent units to obtain the functional unit encoded feature sequence, which contains the time-dependent features of the coupled feature sequence. The coded feature sequences of the functional units are input into the cross-association layer of the association mining network. The association score of any two coded feature sequences of the functional units at each time series position is calculated through a multi-head cross-attention mechanism. The association score is obtained by weighted combination of the attention weight matrix and the coded feature vector. The association scores are arranged according to the functional unit dimension, time series position dimension, and attention head dimension, and input into the tensor output layer of the association mining network. The features are then integrated through a three-dimensional convolutional layer to obtain the association strength tensor.

2. The method as described in claim 1, characterized in that, The step of performing feature coupling processing on the combined sequences in the functional unit data sequence set to generate coupled feature sequences includes: Obtain the carbon emission monitoring subsequence and the associated impact subsequence from the combined sequence, and determine the time series length and feature dimension of the two subsequences. The time series length is the number of positions in the continuous time series, and the feature dimension is the number of feature parameters contained in each time series position. Time series alignment verification is performed on the carbon emission monitoring subsequence and the associated impact subsequence to ensure that the time series lengths of the two subsequences are consistent and that the time series positions correspond one-to-one. The feature vectors of the carbon emission monitoring subsequence at each time series position are interleaved with the feature vectors of the associated influence subsequence at the corresponding time series position. The interleaving is performed by alternating the two feature vectors according to the order of feature parameters to generate an interleaved feature vector. Feature enhancement is performed on the interleaved feature vector by inserting differential features of adjacent time series positions into the interleaved feature vector. The differential features are the difference between the interleaved feature vectors of the current time series position and the previous time series position. The enhanced interleaved feature vectors of all time series positions are arranged in time series order to obtain the coupled feature sequence. The time series length of the coupled feature sequence is the same as that of the original combined sequence, and the feature dimension is a preset multiple of the original feature dimension.

3. The method as described in claim 1, characterized in that, The step of inputting the functional unit encoded feature sequence into the cross-association layer of the association mining network, and calculating the association score of any two functional unit encoded feature sequences at each time series position through a multi-head cross-attention mechanism, includes: Select any two functional unit encoding feature sequences from the functional unit encoding feature sequences as target encoding pairs to obtain the first encoding sequence and the second encoding sequence; The first encoding sequence and the second encoding sequence are respectively input into the query vector generation module and the key-value vector generation module of the multi-head cross-attention mechanism to generate a query vector sequence, a key vector sequence and a value vector sequence. The query vector sequence is obtained by linear transformation of the first encoding sequence, and the key vector sequence and the value vector sequence are obtained by different linear transformations of the second encoding sequence. The query vector sequence, key vector sequence, and value vector sequence are split into multiple header components according to the number of attention heads. Each header component contains a corresponding query sub-vector sequence, key vector sequence, and value vector sequence. Perform a dot product operation on the query subvector sequence and the key subvector sequence for each head component to obtain the original association matrix. The elements of the original association matrix are the dot product results of the query subvector and the key vector. The original correlation matrix is ​​scaled by dividing each element by the square root of the key vector dimension to obtain the scaled correlation matrix. The scaling association matrix is ​​activated to obtain the attention weight matrix, and the elements of the attention weight matrix represent the association weight between the query sub-vector sequence and the key sub-vector sequence. The attention weight matrix and the value vector sequence are weighted and summed to obtain the head-related feature sequence for each head component; The head-related feature sequences of all head components are concatenated dimensionally and integrated through a linear transformation layer to obtain the association score of the target encoding pair at each time series position.

4. The method as described in claim 1, characterized in that, The construction of a time-varying dependency network based on the correlation strength tensor, and the resulting time-varying dependency network parameters, include: The functional unit corresponding to the first dimension of the correlation strength tensor is used as the initial node set of the time-varying dependency network, and each node in the initial node set corresponds to a functional unit. The correlation strength matrix for each time series position is extracted from the correlation strength tensor. The correlation strength matrix is ​​a two-dimensional matrix composed of the first and third dimensions of the correlation strength tensor when the second dimension is fixed. Based on the correlation strength matrix, the connection relationship between nodes at each time series position is determined. For any two nodes, if the corresponding element value in the correlation strength matrix is ​​greater than the preset correlation threshold, a temporary edge connection is established between the nodes at that time series position, and the edge weight is the corresponding element value. The temporary edge connections and edge weights of all time series positions are integrated in time series order to obtain a time-varying edge set, which includes the edge connection states between nodes at different time series positions and the corresponding edge weight values. Extract the feature vector corresponding to each node at each time series position. The node feature vector is obtained by performing element-wise multiplication between the first dimension vector of the node in the association strength tensor and the feature vector of the corresponding time series position of the node's coupling feature sequence. Arrange the node feature vectors in time sequence to obtain the node feature vector time series; Calculate the rate of change of the edge weight of each edge in the time-varying edge set at adjacent time series positions to obtain the edge weight evolution gradient, which is the difference between the edge weight at the current time series position and the edge weight at the previous time series position. By integrating the time series of node feature vectors and the evolution gradient of edge weights, the time-varying network parameters are obtained.

5. The method as described in claim 4, characterized in that, Determining the inter-node connectivity relationships at each time series location based on the correlation strength matrix includes: The correlation strength matrix is ​​subjected to matrix standardization, and the matrix element values ​​are adjusted to a preset range to make the correlation strength matrices at different time series positions comparable. The nonmaximum suppression algorithm is used to process the standardized correlation strength matrix, retaining a preset number of elements with the largest value in each row and setting other elements to zero, thus obtaining a sparse correlation matrix. The sparse correlation matrix is ​​symmetricized. If the element in the i-th row and j-th column of the matrix is ​​non-zero and the element in the j-th row and i-th column is zero, then the element in the j-th row and i-th column is assigned the value of the element in the i-th row and j-th column. Traverse the non-zero elements in the symmetricized sparse incidence matrix and record the row index and column index corresponding to each non-zero element. The row index and column index correspond to the identifiers of the two nodes, respectively. Each non-zero element corresponds to a node identifier pair as a connection between nodes, and the non-zero element value is used as the edge weight of the corresponding connection to obtain the connection relationship between nodes at this time series position.

6. The method as described in claim 4, characterized in that, The extraction of the feature vector corresponding to each node at each time series position includes: Extract the first dimension vector corresponding to the target node from the association strength tensor to obtain the node association feature vector. The dimension of the node association feature vector is consistent with the third dimension of the association strength tensor, and the element value is the degree of association between the target node and other nodes. Extract the coupling feature sub-vectors of the target node at the corresponding time series position from the coupling feature sequence. The coupling feature sub-vectors contain the carbon emission monitoring features and related impact features of the target node. The node-associated feature vector and the coupled feature sub-vector are subjected to dimension adaptation processing. The dimensions of the two vectors are adjusted to the same dimension through linear transformation to obtain the adapted association vector and the adapted coupling vector. Element-wise multiplication is performed on the adaptation association vector and the adaptation coupling vector to obtain the preliminary node feature vector; The preliminary node feature vector is smoothed, and the smoothed preliminary node feature vector is used as the node feature vector of the target node at that time series position.

7. The method as described in claim 1, characterized in that, The step of tracking multi-scale evolution trajectories through the time-varying network parameters to generate a set of multi-scale evolution trajectories includes: A multi-scale time window set is determined, which includes short-term time windows, medium-term time windows, and long-term time windows, and the window length of each time window is a different proportion of the number of positions in the continuous time series; Based on the node feature vector time series in the time-varying network parameters, the node feature vector time series is divided into windows according to a multi-scale time window set to obtain the short-term feature window sequence, medium-term feature window sequence and long-term feature window sequence of each node. The feature window sequence is composed of node feature vectors at multiple consecutive time series positions. Trajectory fitting is performed on the short-term feature window sequence of each node, and the short-term feature evolution trajectory is obtained by a polynomial curve fitting algorithm. The short-term feature evolution trajectory represents the changing trend of the node feature vector within the short-term time window. The same method was used to fit the trajectory of the intermediate feature window sequence and the long-term feature window sequence to obtain the intermediate feature evolution trajectory and the long-term feature evolution trajectory. The gradient values ​​of the edge weights between nodes at each time series position are extracted from the edge weight evolution gradient in the time-varying network parameters, and the short-term gradient window sequence, medium-term gradient window sequence and long-term gradient window sequence are obtained by dividing the time window set according to the multi-scale time window set. Gradient accumulation processing is performed on the short-term gradient window sequence of each node edge, and the sum of gradient values ​​within the window is calculated to obtain the short-term gradient change trajectory. The short-term gradient change trajectory represents the cumulative change of edge weight within the short-term time window. The same method was used to perform gradient accumulation processing on the intermediate gradient window sequence and the long-term gradient window sequence to obtain the intermediate gradient change trajectory and the long-term gradient change trajectory. By integrating the short-term, medium-term, and long-term feature evolution trajectories of all nodes, as well as the short-term, medium-term, and long-term gradient change trajectories of all edges, the multi-scale evolution trajectory set is obtained.

8. The method as described in claim 7, characterized in that, The node feature vector time series based on the time-varying network parameters is divided into windows according to a multi-scale time window set to obtain the short-term feature window sequence, medium-term feature window sequence, and long-term feature window sequence for each node, including: Obtain the total time length of the node feature vector time series, where the total time length is the total number of time series positions; Based on the total time length and the window length of the short-term time window, the sliding step size of the short-term time window is determined. The sliding step size is the number of time sequence positions the window moves, so that there is a preset proportion of overlapping area between adjacent short-term time windows. Starting from the beginning of the node feature vector time series, subsequences are extracted according to the short-term time window length and sliding step size to obtain multiple short-term feature windows. These windows are then arranged in the order of extraction to obtain a short-term feature window sequence. The intermediate sliding step size is determined based on the window length of the intermediate time window and the preset overlap ratio, and the intermediate feature window sequence is obtained by using the same truncation method. The long-term sliding step size is determined based on the window length of the long-term time window and the preset overlap ratio, and the long-term feature window sequence is obtained by using the same truncation method. For each feature window sequence, boundary processing is performed. If the length of the last window is less than the length of the corresponding time window, the window length is padded by copying the feature vector of the last node, so that all windows have the same length.

9. A computer system comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 8.

Images